Ecient Dialogue Strategy to Find Users' Intended Items
from Information Query Results
Kazunori Komatani Tatsuya Kawahara Ryosuke Ito Hiroshi G. Okuno
Graduate School of Informatics, Kyoto University
Kyoto 606-8501, Japan
fkomatani, kawahara, rito, okunog@kuis.kyoto-u.ac.jp
Abstract
Weaddressadialogueframeworkthatnarrows
downtheuser'squeryresultsobtainedbyanin-
formationretrievalsystem. Thefollow-updia-
loguetoconstrainqueryresultsissignicantes-
peciallywiththespeechinterfacessuchastele-
phonesbecausealotofqueryresultscannotbe
presented to the user. The proposed dialogue
frameworkgeneratesguidingquestionsbasedon
an information theoretic criterion to eliminate
retrieved candidates by a spontaneous query
withoutassumingasemanticslotstructure. We
rstdescribeitsconceptongeneralinformation
query tasks, and then deal with a query task
ontheappliancemanualwherestructuredtask
knowledgeisavailable. Ahierarchical conrma-
tion strategyisproposedbymakinguseofatree
structure of the manual, and then three cost
functions for selecting optimal question nodes
arecompared. Experimentalevaluationdemon-
strates that the proposed system helps users
ndtheirintendeditemsmoreeciently.
1 Introduction
In the past years, a great number of spoken
dialogue systems have been developed. Their
typicaltaskdomainsincludeairlineinformation
(Levin et al., 2000;; Potamianos et al., 2000;;
San-Segundoetal.,2000)andtraininformation
(Allenetal.,1996;;Bennacefetal.,1996;;Sturm
etal.,1999;; Lameletal.,1999). Mostofthem
modelspeechunderstandingprocessasconvert-
ingrecognitionresultsintosemanticrepresenta-
tionsequivalenttodatabasequery(SQL)com-
mands,anddialogueprocessasdisambiguating
theirunxedslots. Usually,thesemanticslots
aredenedaprioriandmanually. Theapproach
isworkableonlywhendatastructureoftheap-
plication is well-organized typically as a rela-
tionaldatabase(RDB).
Dierent and more exible approach is
needed for spoken dialogue interfaces to ac-
cessinformationdescribedinlessrigidformat,
in particular normal text database. For the
purpose,informationretrieval(IR)techniqueis
usefultondalistofmatchingdocumentsfrom
the input query. Typically, keywords are ex-
tractedfromthequeryandstatisticalmatching
is performed. Call routing task (Chu-Carroll
and Carpenter, 1998) can be regarded as the
specialcase.
In IR systems, many candidates are usually
obtainedasa queryresult, thusthere isa sig-
nicant problem of how to nd the user's in-
tendeditemamongthem. Especially,eitheron
the telephone or electrical appliances, there is
notalargescreendisplayingthecandidates,and
all the query results cannot bepresented to a
user. Soitisdesirableforthesystemtonarrow
downthequeryresultsinteractively. Moreover,
interactivequeryismorefriendlytonoviceusers
ratherthanrequiringthemtoinputadetailed
queryfromthebeginning.
In this paper, we address a dialogue strat-
egy to nd the user's intended item from the
retrieved result, which is initiated by a spon-
taneous query utterance. In section 2, we de-
scribeamethodtogenerateaguidingquestion
thatnarrowsdownthequeryresultseciently,
using an example of a restaurant query task.
The question is selected based on an informa-
tiontheoreticcriterion. Insection3,wepresent
adialoguemanagementmethodforaquerytask
ontheappliancemanualwherestructuredtask
knowledge is available. We propose a conr-
mationstrategybymakinguseofatreestruc-
tureofthemanual,anddenethreecostfunc-
tionsforselectingquestionnodes. Themethod
isevaluatedbythenumberofaveragedialogue
turns.
Although there are previous studies on
optimizing dialogue strategies (Niimi and
Kobayashi, 1996;; Levin et al., 1997;; Litman
et al., 2000), most of them assume the tasks
of llingsemantic slots that are denitelyand
manually dened, and few focus on follow-up
dialogueofinformationretrieval. Forexample,
(Denecke,1997)proposedamethodtogenerate
guidingquestionsbymakinguseofatreestruc-
ture constructed by unifying retrieved items
based on semantic slots. In this paper, we do
notassumeanystructureofsemanticslots. In-
stead,wemakeuseofdistributionofdocument
statisticsorastructureoftaskknowledge. We
also investigate cost functionsfor optimaldia-
logue control by taking into account of speech
recognitionerrors.
2 Dialogue Strategy in General
Information Query Task
Interactioninaninformationquerytaskcanbe
regarded as a process seeking a common part
between the user's request and system knowl-
edge. In order to help users to nd their in-
tended items from the system knowledge, the
system has to carry out not only interpreting
what users say but also showing the relevant
portionofthesystemknowledgetothem.
We assume that users freely set and retract
querykeys based ontheirpreference for infor-
mationquerysystems. Ifmanycandidatesstill
remainevenafterspecifyingallpossiblehis/her
preferencetothesystem, usersmayhave di-
culty in narrowing down further the query re-
sult. Thus,thesystemshouldgenerateecient
guiding questions to help users nd their in-
tendeditems.
Inthissection,wepresumethesystemknowl-
edgeasapairofanitemandasetofkeywords
(Figure 1). We dene keywords as a set of
words representing contents of the items, and
their categories such as place, food and so on
aregiven. Thisissimilartoindexingwordsin
aconventionalinformationretrievaltask. Note
thatitisnotneededthatthesystemknowledge
isstructuredlikeanRDB.
Keywords are extracted from a user's utter-
ance,andarematchedwiththesystemknowl-
edge. Here, we adopt the following matching
 
RestaurantA
Chinesenoodles,meatdumpling,
Shinjuku,Kabukicho,Ekoda
RestaurantB
Chinesenoodles,meatdumpling,
Shinjuku,Kabukicho,
RestaurantC
Chinesenoodles,meatdumpling,
noodleswithboiled-pork-ribs,
Takadanobaba
RestaurantD
Chinesenoodles,friedgarlic,Yebisu
...
 
Figure1: Anexampleofsystemknowledge
functionforeachitemj.
L
j
=
X
i2K
j

CM
i
log
N
df
i

Here,K
j
isasetofkeywordsforitemj. CM
i
is
a condence measure of speech recognition for
keyword i (KomataniandKawahara,2000), N
isthetotalnumberofitems,anddf
i
isthenum-
ber of items including keyword i. Intuitively,
keywordthatisrecognizedwithhighcondence
anddoesnotappearinmanyitemsgetshigher
likelihoodL
j
by CM
i
and df
i
,respectively.
Then, we deneamount of informationthat
is obtained when the system generates yes/no
questionandtheuseranswersit. Here, C isa
current query condition, A is a condition that
isaddedbythesystem'squestion,andcount(x)
is the numberof items that satisfy the condi-
tion x. The conditionconsists of the conjunc-
tionofthekeywordstheuserspecied. Suppose
each item occurs by equal likelihood, the fol-
lowingequationdenotesthelikelihoodp
0
(A
yes
)
that the yes/no question corresponding to the
addingcondition A willbeansweredas\yes".
p
0
(A
yes
)=
count(C \A)
count(C)
We weight on each item j with the likelihood
L
j
.
p(A
yes
)=
P
j2fC\Ag
L
j
P
j2fCg
L
j
The amount of information that is obtained
when the user's answer is \yes" is represented
asfollows.
I(A
yes
)=log
2
1
p(A
yes
)
The following equation gives H(A), the ex-
pected value of amount of information that is
obtained by generating a question about con-
dition A and getting user's answer (\yes" or
\no").
H(A)=
X
x2fyes;;nog
p(A
x
)log
2
1
p(A
x
)
By calculating H(A) for all conditions A that
can be added to the current query condition,
thesystemgeneratesthequestionthathasthe
maximumvalueof H(A). Thequestionisgen-
erated using the category information of each
keyword.
Becausetheobtainedcondition A isselected
by a viewpointof narrowingdown the current
set of items eciently, the selected condition
maybeunimportantfortheuser. Insuchacase,
it isnot cooperative to force the user an ar-
mativeornegativereply. Oursystemdoesnot
forcethereluctantdecisionbyallowingtheuser
to say \It doesnot matter anyhow.". Instead,
thesystempresentsthesecondbestproposal.
Weexplainthemethodwiththefollowingex-
ample in our restaurant query system in the
Tokyoarea.Whenausersays,\Pleasetellmea
restaurantwhereIcaneatChinesenoodleand
meat dumpling in Shinjuku area.", three key-
wordsareextracted: \Shinjuku",\Chinesenoo-
dle"and \meat dumpling". Asa resultof the
matchingusingthesethreekeywords,11query
results are obtained. It is not cooperative to
readoutallofthe11queryresultswithaTTS
(text-to-speech)system. Here,theexpectedval-
ues of amount of information H(A) are calcu-
latedforeachconditionthatcorrespondstokey-
wordsincludedinthematcheditemsexceptfor
thethreekeywords,\Shinjuku",\Chinesenoo-
dle"and\meatdumpling". Then,weselectthe
keyword \noodles with boiled-pork-ribs" that
hasthemaximumvalueH(A). Bygeneratinga
questionlike\Wouldyoulikeonewhichserves
noodleswithboiled-pork-ribs?",andobtaining
areplyfromtheuser,thesystemaddsthenew
conditionandnarrowsdownthecandidatesef-
ciently. If the user thinks that the condition
\noodles with boiled-pork-ribs" is not impor-
tant and tells the system so (for example \Ei-
therwilldo."),thesystemcanshowthesecond
best proposal, \Would you like one located in
Kabukichoarea?". Thus, the queryresultcan
benarroweddownwithoutforcingtheuserun-
naturalyes/noanswers.
3 Dialogue Strategy for Query on
Appliance Manuals
Inthissection,wepresentanotherecientsolu-
tioninthecasethatthestructureorhierarchy
of task knowledge is available. The task here
is to nd the appropriate item in the manual
ofelectricapplianceswithaspokendialoguein-
terface. Suchaninterfacewillbeusefulasthe
recent appliances become complex with many
features and so are their manuals. In the ap-
pliances such as VTR (Video Tape Recorder)
andFAX machines, thereisnota largescreen
todisplaythelistofmatchedcandidatestobe
selected by the user. Therefore, we address a
spokendialoguestrategytodeterminethemost
appropriateonefromthelistofcandidates.
Analternativesystemdesignistheuseofdi-
rectory search, as adopted in voice portal sys-
tems, where the documents are hierarchically
structuredandthesystempromptsuserstose-
lect one of the menu from the top to the leaf.
Themethodisrigidandnotuser-friendlysince
users often have trouble in selection and want
to specify by their own expression. The pro-
posedsystemallowsuserstomakequeriesspon-
taneouslyandmakesuseofthedirectorystruc-
tureinthefollow-updialoguetodeterminethe
mostappropriateone.
3.1 System Overview
AnoverviewofthesystemisillustratedinFig-
ure2. Itconsistsoffollowingprocesses.
1. Keywordspottingfromuserutterancesus-
inganASR(automaticspeechrecognition)
system(Kawaharaetal.,1998)
A natural spoken language query is ac-
ceptedandkeywordsareextracted. Acon-
dence measure CM
i
is assigned to each
keyword i basedontheN-bestrecognition
result(KomataniandKawahara,2000).
yes/no
system user
manual
tree
structure
entries
keyword spotting
matching
follow-up
dialogue
keywords with
confidence
entries with
likelihood
result
spoken query
Figure2: Systemoverview
2. Matchingwithmanualitems(documents)
Theextractedkeywordsarematchedwith
a set of manual items. The matching is
performedontheinitialportion(indexand
rst summary paragraph) of each manual
section. We adopt thefollowingmatching
scorefunctionforanitem j. K
j
isasetof
keywordsforitem j.
L
j
=
1
n
j
X
i2K
j
(CM
i
log
N
df
i
)
Here, df
i
isthenumberofitemsthatcon-
tainkeyword i referredasadocumentfre-
quencyandN isthetotalnumberofitems.
The inverse document frequency (idf) is
weighted with a condence measure CM
i
andsummedover keywords, then normal-
izedby n
j
,thenumberofkeywordsinthe
item j.
3. Generatingdialoguetodeterminethemost
appropriateonefromthelistofcandidates
As a result of the matching, many candi-
datesareusuallyfound. Theymayinclude
irrelevant ones because of speech recogni-
tionerrors. Butitisnotpracticaltoread
outalloftheminorderwithaTTS(text-
to-speech) system. Therefore, dialogue is
invoked to narrow down to the intended
one. Thisdialogueisrestrictedtosystem-
initiated \yes/no" questions in order to
play record search setting
normal
play
slow
play
.............
Figure3: Exampleoftreestructureofmanual
avoid furtherrecognitionerrors and back-
up dialogue. The dialogue strategy is ex-
plainedinthenextsubsection.
3.2 Dialogue Strategy using Structure
of Manual
Ifoneofthecandidatesismoreplausiblethan
others with a signicant margin, we should
makeconrmationonit. Whentherearemany
candidates with similar condence and they
canbehierarchicallygroupedintoseveralcate-
gories, we hadbetter rstidentifywhichcate-
gorytheintendedonebelongsto. Inthiswork,
wemakeuseofthesectionstructureoftheman-
ual,i.e. sectionistherstlayer,sub-sectionis
thesecond-layer,andsoon. Thetreestructure
is automatically derived from its table of con-
tents. AnexampleforVTRmanualisshownin
Figure3.
For each node of the tree, likelihood L
0
j
is
assignedasfollows.
 For a leaf node, the matching score L
j
is
assignedafternormalizingsothatthesum
overallleaves(manualitems)is1.0.
 Fora non-leafnode, thesumofthelikeli-
hoodofitschildrennodesisassigned.
Then,adialogueisgeneratedasfollows.
1. Among ancestor nodes of the leaf of the
largestlikelihoodL
0
j
,pickuptheonewhose
heuristic cost function described below is
smallest.
2. Makea\yes/no"questiononthenode,for
example\Doyouwanttoknowabout...?".
The content of the question is associated
withthesectiontitle.
3. Iftheuser'sansweris\yes",eliminatethe
nodes other than descendants of the con-
0.4 0.1 0 0.1 0.2 0.2 0 0
0.5 0.1 0.4 0
0.6 0.4
0.40.10 00000
0.5 0 0 0
0.5 0
0 0 0 0.1 0.2 0.2 0 0
0
0.1 0.4 0
0.1 0.4
"Yes" "No"
leaf with
best score
selected by
cost function
generate a "yes-no" question
on selected node
Figure4: Useofmanualstructureandcostfunctionfordialoguecontrol
rmednode. Ifthe answer is\no", elimi-
natealldescendantsofthedeniednode.
4. Repeattheprocessuntilonlyonenode(or
lessthanathreshold )remains.
TheaboveprocessesareillustratedinFigure4.
Wedenefollowingthreeheuristiccostfunc-
tionsinordertorealizeanecientdialogue.
 h
1
(j)=jL
0
j
;0:5j
Thismakesaquestiononthemostambigu-
ousnodewhoselikelihoodL
0
j
interpretedas
aposterioriprobabilityiscloseto0.5.
 h
2
(j) = L
0
j
 Node
j
(yes) + (1 ; L
0
j
) 
Node
j
(no)
Here, Node
j
is the number of remaining
nodes when the answer is \yes" or \no".
Thisfunctiontakestheapproximatenum-
beroffollowingquestionsintoaccount.
 h
3
(j) = L
0
j
 Ques
j
(yes) + (1 ; L
0
j
) 
Ques
j
(no)+1
Ques
j
istheestimatednumberoftimesof
questionsneededwhentheansweris\yes"
or\no". Itiscomputedrecursivelybyex-
pandingthesub-tree,andisassignedwith
0 when the numberof remaining nodes is
under a threshold (). This  means the
numberofcandidatesthatcanbepresented
tousers. Here,weset=3.
Theseareexperimentallycomparedinthenext
subsection.
3.3 Experimental Evaluation
3.3.1 Task and System Implementation
The proposed system is implemented for the
query task on a VTR manual that consists of
111pagesand47items. Thederivedtreestruc-
tureisofthreelevels. Thenumberofkeywords
usedformatchingis137.
Thespeechrecognitionsystemisbasedonour
largevocabularycontinuousspeechrecognition
engine Julius (Lee et al., 2001). The language
modelisinitiallybasedonanitestate gram-
marandextendedtocombinestatisticalmodels
derived from the domain-specic corpus (Ko-
matani et al., 2001), that is the manual text
in this task. The acoustic model is a gender-
dependent phonetic tied-mixture (PTM) tri-
phonemodel(Leeetal.,2000)trainedwiththe
40-hourJNASspeechcorpus.
Forcollectingevaluationdata,wehad14sub-
jectsandeachmade10queriesongivenscenar-
ios(querysentencesarenotgiven),andseveral
spontaneous querieswithoutany scenarios. In
total, we had 195 query utterances, of which
Table1: Evaluationresultwithtextinput
#matchedcandidates 12.4
querysuccessrate 93%
averagerankofcorrectitem
(#turnsbybaseline)
3.2
#turnsby h
1
h
2
h
3
proposedcostfunctions 2.4 2.5 2.8
157couldbecopedwiththegivenmanual,thus
usedasthetest-set. Samplequeriesare\Iwant
tochangetherecordingreservation." and\Can
IwatchTVwhilerecordinganotherprogram?"
Asforevaluationmeasures,werstcompute
therateofquerysuccesswherethecorrectman-
ualitemiscontainedinthecandidatelistbythe
initialmatching. Then,thesystemisevaluated
by the necessary dialogue turns equivalent to
thenumberofquestionsbeforethecorrectitem
is identied. It is compared with the baseline
casewherethecandidatesarepresentedtothe
userinorderofthematchingscore L
j
andthe
number of dialogue turns is equivalent to the
rankofthecorrectitem.
3.3.2 Evaluation with Text Input
Atrst,thesystemisevaluatedwithtextinput,
which is transcription of the collected queries.
TheresultisshowninTable1.
Ontheaverage, thematchingresultconsists
of12.4candidatesandcontainscorrectonefor
93%ofthetractablequeries. Theaveragerank
of the correct item is 3.2, which means, if we
make conrmation in order of the matching
score L
j
, we need 3.2 turns on the average.
Withdialoguebasedontheheuristiccostfunc-
tions,itcanbereducedto2.4(h
1
),2.5(h
2
)and
2.8(h
3
),respectively.
We have not yet identied the reason why
performance by the apparently most accurate
function h
3
is not good. We conjuncture that
the dierence of the cost functions does not
matter so much in this framework as long as
theyarereasonable.
3.3.3 Evaluation with Speech Input
Next, we made experiments using the spoken
queriesandthespeechrecognitionsystem. The
distributionof recognized keywords and corre-
spondingcondencemeasuresisshowninTable
2. Theprecisionforthekeywordswithhighcon-
Table3: Evaluationresultwithspeechinput
#matchedcandidates 13.3
querysuccessrate 87%
averagerankofcorrectitem
(#turnsbybaseline)
4.1
#turnsby h
1
h
2
h
3
proposedcostfunctions 2.9 2.9 3.2
dencemeasuresisbetter, thusthecondence
measure works well. Summaryof the result is
giveninTable3.
Theaveragenumberofmatcheditemsis13.3
andthesuccessrateis87%. Somedegradation
from the case of text input is observed. The
averagerankofthecorrectitemis4.1. Forref-
erence,ifwedonotusethecondencemeasure
CM
i
,thegureis4.4,whichveriestheeectof
thecondencemeasure. Theproposeddialogue
strategy with either heuristic function reaches
thecorrectoneinaround3turns,whichis30%
reductioncomparedwiththebaseline.
Itshouldbenoticedthat,althoughtheinitial
matching accuracy is lowered with the speech
input,theimprovementbytheproposedstrat-
egyislargerandthenumberofdialogueturns
isclosetothetext-inputcase. Theresultcon-
rms that the proposed framework is eective
inspeechinterface.
4 Conclusion
We present a method to generate guiding ut-
terances for narrowing down users' query re-
sultsobtained by an informationretrieval sys-
tem. Byselectingthemostecientitem,thedi-
alogueisrestrictedtosystem-initiated\yes/no"
questions. Wehaveevaluatedourmethodwith
a query task on the appliance manual where
structured task knowledge is available. The
numberofaveragedialogueturnsisreducedby
about30%comparedwithabaselinemethodin
whichthecandidatesareconrmedaccordingto
theirmatchingscores. Thisresultdemonstrates
thattheproposedsystemhelpsusersndtheir
intendeditemsmoreeciently.
Table2: Theprecisionofkeywordsandtheircondencemeasures
condencemeasureofkeyword 1 1-0.9 0.9-0.8 0.8-0.7 0.7- total
#correctlyrecognizedwords 279 15 10 18 16 338
#incorrectlyrecognizedwords 63 17 20 49 60 209
precision 82% 47% 33% 27% 21% 62%

References

J.F. Allen, B.W. Miller, E.K. Ringger, and
T.Sikorski. 1996. Arobustsystemfornatu-
ralspokendialogue. In Proc. of the 34th An-
nual Meeting of the Association for Compu-
tational Linguistics (ACL-96),pages62{70.

S. Bennacef, L. Devillers, S. Rosset, and
L. Lamel. 1996. Dialog in the RAILTEL
telephone-basedsystem. In Proc. Int'l Conf.
on Spoken Language Processing.

JenniferChu-CarrollandBobCarpenter. 1998.
Dialogue management in vector-based call
routing. In Proc. of COLING-ACL98,pages
256{262.

MatthiasDenecke. 1997. Aninformation-based
approach for guiding multi-modal human-
computer-interaction. InProc. of the 15th In-
ternational Joint ConferenceonArticial In-
telligence (IJCAI-97).

T. Kawahara, C.-H. Lee, and B.-H. Juang.
1998. Flexible speech understanding based
on combined key-phrase detection and veri-
cation. IEEE Trans. on Speech and Audio
Processing,6(6):558{568.

K.KomataniandT.Kawahara. 2000. Flexible
mixed-initiative dialogue management using
concept-level condence measures of speech
recognizeroutput. In Proc. Int'l Conf. Com-
putational Linguistics (COLING),pages467{
473.

K. Komatani, K. Tanaka, H. Kashima, and
T. Kawahara. 2001. Domain-independent
spoken dialogue platform using key-phrase
spottingbasedoncombinedlanguagemodel.
In Proc. European Conf. Speech Commun. &
Tech. (EUROSPEECH),pages1319{1322.

L.F.Lamel,S.Rosset,J-L.S.Gauvain,andS.K.
Bennacef. 1999. The LIMSI ARISE system
fortraintravelinformation. In Proc. of Int'l
Conf. on Acustics, Speech and Signal Process-
ing (ICASSP).

A. Lee, T. Kawahara, K. Takeda, and
K. Shikano. 2000. A new phonetic tied-
mixturemodelforecientdecoding. InProc.
of Int'l Conf. on Acustics, Speech and Signal
Processing (ICASSP),pages1269{1272.

A. Lee, T. Kawahara, and K. Shikano. 2001.
Julius { an open source real-time large vo-
cabulary recognition engine. In Proc. Euro-
pean Conf. Speech Commun. & Tech. (EU-
ROSPEECH),pages1691{1694.

E. Levin, R. Pieraccini, and W. Eckert.
1997. Learningdialoguestrategieswithinthe
markovdecisionprocessframework. In Proc.
IEEE Workshop on Automatic Speech Recog-
nition and Understanding,pages72{79.

E.Levin,S.Narayanan, R.Pieraccini,K.Bia-
tov, E. Bocchieri, G. Di Fabbrizio, W. Eck-
ert,S.Lee,A.Pokrovsky,M.Rahim,P.Rus-
citti, and M. Walker. 2000. The AT&T-
DARPA communicator mixed-initiativespo-
kendialoguesystem. In Proc. Int'l Conf. on
Spoken Language Processing.

Diane J. Litman, Michael S. Kearns, Satinder
Singh, andMarilynA. Walker. 2000. Auto-
maticoptimizationofdialoguemanagement.
In Proc. Int'l Conf. Computational Linguis-
tics (COLING),pages502{508.

Y.NiimiandY.Kobayashi. 1996. Adialogcon-
trolstrategybasedonthereliabilityofspeech
recognition. In Proc. Int'l Conf. on Spoken
Language Processing.

AlexandrosPotamianos,EgbertAmmicht,and
Hong-Kwang J. Kuo. 2000. Dialogue man-
agement in the bell labs communicator sys-
tem. InProc. Int'l Conf. on Spoken Language
Processing.

R. San-Segundo, B. Pellom, W. Ward, and
J.Pardo. 2000. Condencemeasuresfordia-
loguemanagement inthe CUcommunicator
system. In Proc. of Int'l Conf. on Acustics,
Speech and Signal Processing (ICASSP).

J.Sturm,E.Os,andL.Boves. 1999. Issuesin
spoken dialogue systems: Experiences with
the Dutch ARISE system. In Proc. ESCA
workshop on Interactive Dialogue in Multi-
Modal Systems.
